Twelve days ago I said the fix for Ainara would be leaving SDXL behind. That writeup (https://civitai.red/articles/33842/what-i-fixed-what-i-did-not-and-the-one-rule-about-captions-that-is-wrong) ended with a list of things I'd do differently for V2: vary the background, vary the clothing, get real full body shots, and stop training on an architecture built before I even had this account. Here's what happened when I actually did it.
The dataset got bigger, not just different
47 images this time, up from 25, pulled from a personal photo archive that turned out to be the same shoots exported into half a dozen overlapping folders and zip files at different resolutions. Most of the "new" images were duplicates in disguise. I hashed everything and kept one copy of each -- the real number of unique shots was smaller than it looked, but wider in range: full body, close up, different backgrounds, different outfits. Exactly what V1.1 was missing.
I found the same bug again
The last article's headline lesson was captions describing a different image than the one they were attached to. I said I'd fixed it. Building this dataset, I found it again -- worse. One caption from the old batch claimed "no flowers no jewelry" on a photo that clearly showed a flower crown and a choker. I patched that one, spot-checked a second file, and called it done. It wasn't. A full pass turned up roughly a dozen more mismatches in one subset alone. Every caption in the set got rewritten from the actual image, not the old text. Same failure mode as before, just further along before I caught it.
Krea 2 forces the format question
Krea 2's trainer wants tags, not sentences -- comma separated phrases, not prose. That's a bigger rewrite than it sounds: every caption switched style at the same time the content got fixed, on a dataset three times bigger than V1.1's. Trigger word stayed ainara.
The gotcha nobody warns you about
Fourteen of the 47 images were 1120x2240 -- over the trainer's 2048px cap, silently. No error until upload. Resized down, aspect ratio preserved, moved on. If you're building a set from mixed-resolution exports, check this before you upload, not after.
Krea 2 over Z-Image Turbo, and why
Z-Image Turbo is cheaper (1000 Buzz against Krea 2's 1400, same 2000 steps / 10 checkpoints) and API-scriptable, Krea 2 is UI-only. I went with Krea 2 anyway -- it's the model that wins on identity preservation in the research I could find, and for a character LoRA that's the number that matters, not the price.
Does it actually work?
I compared all 10 training checkpoints against the same three test prompts before picking one. Identity held from checkpoint 1 through 10, no drift, no collapse. At full resolution the skin finally looks like skin -- pores, texture, none of the plastic-doll flatness earlier SDXL attempts kept producing. Went with the recommended final checkpoint.
Full body is still untested past the training samples. Same offer as last time: if you get a strength or sampler combination that holds up in a wide shot, tell me, I'll credit you in the description.
Ainara V2 is up. https://civitai.red/models/2886424/ainara
Trigger word: ainara.